Many automated approaches have been proposed in literature to quantify clinically relevant wound features based on image processing analysis, aiming at removing human subjectivity and accelerate clinical practice. In this work we present a fully automated image processing pipeline leveraging deep learning and a large wound segmentation dataset to perform wound detection and following prediction of the Photographic Wound Assessment Tool (PWAT), automatizing the clinical judgement of the adequate wound healing. Starting from images acquired by smartphone cameras, a series of textural and morphological features are extracted from the wound areas, aiming to mimic the typical clinical considerations for wound assessment. The resulting extracted features can be easily interpreted by the clinician and allow a quantitative estimation of the PWAT scores. The features extracted from the region-of-interests detected by our pre-trained neural network model correctly predict the PWAT scale values with a Spearman's correlation coefficient of 0.85 on a set of unseen images. The obtained results agree with the current state-of-the-art and provide a benchmark for future artificial intelligence applications in this research field.

Background: Radiomics is a field of research medicine and data science in which quantitative imaging features are extracted from medical images and successively analyzed to develop models for providing diagnostic, prognostic, and predictive information. The purpose of this work was to develop a machine learning model to predict the survival probability of 85 cervical cancer patients using PET and CT radiomic features as predictors. Methods: Initially, the patients were divided into two mutually exclusive sets: a training set containing 80% of the data and a testing set containing the remaining 20%. The entire analysis was separately conducted for CT and PET features. Genetic algorithms and LASSO regression were used to perform feature selection on the initial PET and CT feature sets. Two different survival models were employed: the Cox proportional hazard model and random survival forest. The Cox model was built using the subset of features obtained with the feature selection process, while all the available features were used for the random survival forest model. The models were trained on the training set; cross-validation was used to fine-tune the models and to obtain a preliminary measurement of the performance. The models were then validated on the test set, using the concordance index as the metric. In addition, alternative versions of the models were developed using tumor recurrence as an adjunct feature to evaluate its impact on predictive performance. Finally, the selected CT and PET features were combined to build a further Cox model. Results: The genetic algorithm was superior to the LASSO regression for feature selection. The best performing model was the Cox model, which was built using the selected CT features; it achieved a concordance index score of 0.707. With the addition of tumor recurrence as a predictive feature, the Cox CT model reached a concordance index score of 0.776. PET features, however, proved to be inadequate for survival prediction. The CT model performed better than the model with combined PET and CT features. Conclusions: The results showed that radiomic features can be used to successfully predict survival probability in cervical cancer patients. In particular, CT radiomic features proved to be better predictors than PET radiomic features in this specific case.

Background: Benign renal tumors, such as renal oncocytoma (RO), can be erroneously diagnosed as malignant renal cell carcinomas (RCC), because of their similar imaging features. Computer-aided systems leveraging radiomic features can be used to better discriminate benign renal tumors from the malignant ones. The purpose of this work was to build a machine learning model to distinguish RO from clear cell RCC (ccRCC). Method: We collected CT images of 77 patients, with 30 cases of RO (39%) and 47 cases of ccRCC (61%). Radiomic features were extracted both from the tumor volumes identified by the clinicians and from the tumor’s zone of transition (ZOT). We used a genetic algorithm to perform feature selection, identifying the most descriptive set of features for the tumor classification. We built a decision tree classifier to distinguish between ROs and ccRCCs. We proposed two versions of the pipeline: in the first one, the feature selection was performed before the splitting of the data, while in the second one, the feature selection was performed after, i.e., on the training data only. We evaluated the efficiency of the two pipelines in cancer classification. Results: The ZOT features were found to be the most predictive by the genetic algorithm. The pipeline with the feature selection performed on the whole dataset obtained an average ROC AUC score of 0.87 ± 0.09. The second pipeline, in which the feature selection was performed on the training data only, obtained an average ROC AUC score of 0.62 ± 0.17. Conclusions: The obtained results confirm the efficiency of ZOT radiomic features in capturing the renal tumor characteristics. We showed that there is a significant difference in the performances of the two proposed pipelines, highlighting how some already published radiomic analyses could be too optimistic about the real generalization capabilities of the models.

Aim: Machine learning (ML) and deep learning (DL) predictive models have been employed widely in clinical settings. Their potential support and aid to the clinician of providing an objective measure that can be shared among different centers enables the possibility of building more robust multicentric studies. This study aimed to propose a user-friendly and low-cost tool for COVID-19 mortality prediction using both an ML and a DL approach. Method: We enrolled 2348 patients from several hospitals in the Province of Reggio Emilia. Overall, 19 clinical features were provided by the Radiology Units of Azienda USL-IRCCS of Reggio Emilia, and 5892 radiomic features were extracted from each COVID-19 patient’s high-resolution computed tomography. We built and trained two classifiers to predict COVID-19 mortality: a machine learning algorithm, or support vector machine (SVM), and a deep learning model, or feedforward neural network (FNN). In order to evaluate the impact of the different feature sets on the final performance of the classifiers, we repeated the training session three times, first using only clinical features, then employing only radiomic features, and finally combining both information. Results: We obtained similar performances for both the machine learning and deep learning algorithms, with the best area under the receiver operating characteristic (ROC) curve, or AUC, obtained exploiting both clinical and radiomic information: 0.803 for the machine learning model and 0.864 for the deep learning model. Conclusions: Our work, performed on large and heterogeneous datasets (i.e., data from different CT scanners), confirms the results obtained in the recent literature. Such algorithms have the potential to be included in a clinical practice framework since they can not only be applied to COVID-19 mortality prediction but also to other classification problems such as diabetic prediction, asthma prediction, and cancer metastases prediction. Our study proves that the lesion’s inhomogeneity depicted by radiomic features combined with clinical information is relevant for COVID-19 mortality prediction.

Background: Microvascular invasion (MVI) is a necessary step in the metastatic evolution of hepatocellular carcinoma liver tumors. Predicting the onset of MVI in the initial stages of the tumors could improve patient survival and the quality of life. In this study, the possibility of using radiomic features to predict the presence/absence of MVI was evaluated. Methods: Multiphase contrast-enhanced computed tomography (CECT) images were collected from 49 patients, and the radiomic features were extracted from the tumor region and the zone of transition. The most-relevant features were selected; the dataset was balanced, and the presence/absence of MVI was classified. The dataset was split into training and test sets in three ways using cross-validation: the first applied feature selection and dataset balancing outside cross-validation; the second applied dataset balancing outside and feature selection inside; the third applied the entire pipeline inside the cross-validation procedure. Results: The features from the tumor areas on CECT showed both the portal and the arterial phases to be the most predictive. The three pipelines showed receiver operating characteristic area under the curve (ROC AUC) scores of 0.89, 0.84, and 0.61, respectively. Conclusions: The results obtained confirmed the efficiency of multiphase CECT and the ZOT in detecting MVI. The results showed a significant difference in the performance of the three pipelines, highlighting that a non-rigorous pipeline design could lead to model performance and generalization capabilities that are too optimistic.

Wound management is a fundamental task in standard clinical practice. Automated solutions already exist for humans, but there is a lack of applications regarding wound management for pets. Precise and efficient wound assessment is helpful to improve diagnosis and to increase the effectiveness of treatment plans for chronic wounds. In this work, we introduced a novel pipeline for the segmentation of pet wound images. Starting from a model pre-trained on human-based wound images, we applied a combination of transfer learning (TL) and active semi-supervised learning (ASSL) to automatically label a large dataset. Additionally, we provided a guideline for future applications of TL+ASSL training strategy on image datasets. We compared the effectiveness of the proposed training strategy, monitoring the performance of an EfficientNet-b3 U-Net model against the lighter solution provided by a MobileNet-v2 U-Net model. We obtained 80% of correctly segmented images after five rounds of ASSL training. The EfficientNet-b3 U-Net model significantly outperformed the MobileNet-v2 one. We proved that the number of available samples is a key factor for the correct usage of ASSL training. The proposed approach is a viable solution to reduce the time required for the generation of a segmentation dataset.